Mini toolsRobostral Navigate
By Mistral AI
Model family: Mistral
Robostral Navigate is an 8B parameter embodied navigation model built entirely in house, without relying on existing open source vision language models. It is initialized from a vision language model specialized in pointing, counting, and object localization, then trained on about 400,000 simulated trajectories across 6,000 scenes using a token efficient, prefix caching supervised training method. Performance is further improved with online reinforcement learning using the CISPO algorithm, raising success rate by 3.2 percent. The model takes RGB images and a plain language instruction and predicts the next navigation target through image based pointing, or local coordinate displacements when the target is out of view. It runs on wheeled, legged, and flying robots, generalizes across robot types and camera intrinsics, and targets navigation tasks in offices, residential and commercial buildings, and outdoor settings.
New
Multimodal
Overview
Robostral Navigate is an 8B parameter robotics navigation model that uses a single RGB camera and natural language instructions to autonomously guide robots through complex environments. It predicts the next target location via image based pointing and falls back to local coordinate displacements when the target is out of view, achieving 76.6 percent success on R2R-CE validation unseen without depth sensors, LiDAR, or multiple cameras.
About Mistral AI
Mistral AI is a company that specializes in artificial intelligence and machine learning solutions.
View Company Profile
Last updated: July 8, 2026
